`claude-api` skill eagerly injects ~156k tokens, exhausting the context window in a single tool call

# `claude-api` skill eagerly injects ~156k tokens, exhausting the context window in a single tool call

## Environment

| | |
|---|---|
| Claude Code | 2.1.220 |
| Skill | `claude-api` (bundled at `%LOCALAPPDATA%\Temp\claude\bundled-skills\2.1.220\\claude-api`; also published via `anthropic-agent-skills` → `github.com/anthropics/skills`) |
| Model | `claude-opus-5` (default context) |
| OS | Windows 11 (10.0.22631), Git Bash |
| Node | v24.14.0 |

## Summary

Invoking `Skill(claude-api)` injects a single message of **624,401 characters (~156k tokens)** — the concatenation of `SKILL.md` plus **31 reference documents**. On a default-context model this consumes the entire window, and the very next API call fails with `Prompt is too long`. The session is unrecoverable.

This is not a slow degradation. It is one tool call from a fresh session to a dead session.

## Reproduction

1. Start a new session on `claude-opus-5` (default context).
2. Prompt with a request to edit a Markdown file that mentions model names in prose, e.g.:

> Opus 5 is now available and is more capable than Fable 5. Please update the model list in our global CLAUDE.md accordingly.

3. The model reads `~/.claude/CLAUDE.md` (~1 KB), then invokes `Skill(claude-api)` per the skill's trigger.
4. Next assistant turn: `Prompt is too long`.

The task is editing a Markdown table in a config file. No code, no API calls, no SDK usage anywhere in scope.

*(The prompt above is minimized from the one that produced the transcript below; it retains the tokens that match the trigger — `Opus`, `Fable`, `CLAUDE.md`.)*

### Transcript evidence

| Entry | Type | Size |
|---|---|---|
| 6 | user prompt | 683 B |
| 13 | `Read` of a ~1 KB Markdown file | ~1 KB |
| 14 | `tool_use` / `Skill` | 1.2 KB |
| **17** | **skill payload** | **624,401 chars (~156k tokens)** |
| 19 | assistant | `Prompt is too long` |

## Expected vs. actual

**Expected:** the skill loads its entry document and reads references on demand, as the progressive-disclosure model for skills implies.

**Actual:** 32 documents arrive at once, before the model knows which — if any — are relevant.

## Analysis

### 1. Over half the payload is provably irrelevant to any given task

Composition of the 624,401 characters:

| Segment | Chars | Share |
|---|---|---|
| `shared/model-migration.md` | 143,685 | 23% |
| Managed Agents suite (14 files) | 202,069 | 32% |
| `SKILL.md` | 68,375 | 11% |
| Everything else (17 files) | 210,272 | 34% |

**`shared/model-migration.md` is the single largest file at 23% of the payload — and `SKILL.md`'s own subcommand table says to load it conditionally:**

> | `migrate` | … **Read `shared/model-migration.md` immediately** … |

It is gated behind an explicit subcommand, then loaded eagerly regardless. The Managed Agents suite (another 32%) is likewise a distinct feature area that a config-file edit never touches. **55% of the payload is two on-demand feature areas.**

Language detection does work — only TypeScript variants were included, not all nine languages. The eager loading of `shared/` defeats the savings.

### 2. No lighter invocation exists

The skill exposes exactly one subcommand, `migrate`. It cannot reduce the payload, because the dispatch instruction operates on already-loaded text:

> If the User Request at the bottom of this prompt is a bare subcommand string (no prose), **search every Subcommands table in this document** …

The document must be fully injected before the subcommand is read. There is no argument that loads less.

### 3. The trigger is far too broad for the payload size

From the skill description:

> TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`) …

Any prompt containing the bare string "Opus", "Sonnet", or "Claude" matches — including editing a file *named* `CLAUDE.md`. Coupled with a 156k-token payload and language that explicitly preempts the model's judgement (*"don't skip because it looks like a one-liner"*), the failure mode is systematic rather than incidental.

### 4. The payload was stale for the task that summoned it

The bundled content is stamped `## Current Models (cached: 2026-06-24)` and describes **Fable 5** as *"most capable widely released model"*. It contains no knowledge of Opus 5 — the exact fact being added to the file. The skill exhausted the session to supply information it did not have.

## Suggested fixes

Roughly in order of impact:

1. **Make `shared/` progressive.** Ship `SKILL.md` with an index of references and let the model `Read` what it needs. This alone cuts ~89% of the payload.
2. **Honour the existing conditional gates.** `model-migration.md` and the Managed Agents suite are already documented as on-demand; load them that way.
3. **Narrow the trigger.** Require an API/SDK signal (an `anthropic` import, an API call site, a model-id string in code) rather than any mention of a model name in prose. Editing a `.md` file should not match.
4. **Soften the anti-judgement language.** *"Don't skip because it looks like a one-liner"* suppresses exactly the check that would have avoided this.
5. **Budget-aware loading.** A skill about to consume ~78% of a 200k window should degrade to its index rather than fail the session.

## Impact

Any user on a default-context model who mentions a Claude model name loses their session. It is most likely to fire on trivial tasks — config edits, doc updates, casual questions about model choice — where the cost is least justified and the user has least warning.

**Workaround:** `"permissions": { "deny": ["Skill(claude-api)"] }` in `settings.json`, or run in a 1M-context session.

Start by inspecting SKILL.md's trigger and subcommand/reference tables, then compare the eager payload with shared/model-migration.md and the Managed Agents files. Reproduce the claude-api invocation from the issue and verify that a Markdown-only edit no longer injects all 32 documents or causes the next API call to fail with "Prompt is too long".

Do mô hình lập chỉ mục viết ra từ nội dung của issue.